Introduction

1. Introduction

1.1 Abstract

This document defines the M2M (Machine-to-Machine) Protocol, a token-optimized compression scheme for Large Language Model (LLM) API traffic. Unlike traditional compression algorithms that reduce bytes but increase token count due to Base64 encoding, M2M Protocol achieves 25-40% token reduction through semantic key abbreviation, value substitution, and default parameter elimination.

This specification defines the wire format, compression mappings, session negotiation, and security considerations for M2M Protocol version 1.0.

1.2 Background

1.2.1 The Problem

LLM APIs charge based on token count, not bytes. Traditional compression algorithms (gzip, brotli, zstd) reduce byte size but produce binary output requiring Base64 encoding, which typically increases token count by 33%.

Example with gzip:

Original:     68 bytes  → 42 tokens
Gzip+Base64:  52 bytes  → 58 tokens (+38% tokens)

1.2.2 The Solution

M2M Protocol applies semantic compression that preserves JSON structure while reducing both bytes and tokens:

Original:     68 bytes  → 42 tokens
M2M Token:    45 bytes  → 29 tokens (-31% tokens)

Key techniques:

1. Key abbreviation: "messages" → "m"
2. Value substitution: "assistant" → "a"
3. Model abbreviation: "gpt-4o" → "4o"
4. Default elimination: Remove "temperature": 1.0 (default)

1.3 Protocol Overview

M2M Protocol operates in two modes:

1.3.1 Stateless Mode

Direct compression/decompression without session establishment:

Client                          Server
   |                               |
   |--- Compressed Request ------->|
   |<-- Compressed Response -------|

1.3.2 Session Mode

Full protocol with capability negotiation:

Client                          Server
   |                               |
   |-------- HELLO --------------->|
   |<------- ACCEPT ---------------|
   |                               |
   |======= DATA (compressed) ====>|
   |<====== DATA (compressed) =====|
   |                               |
   |-------- CLOSE --------------->|

1.4 Design Goals

1. Token Reduction: Optimize for LLM tokenizer output, not just bytes
2. JSON Compatibility: Compressed output is valid JSON
3. Low Latency: Sub-millisecond compression overhead
4. Backward Compatibility: Graceful fallback to uncompressed
5. Extensibility: Support for new compression algorithms
6. Security: Optional threat detection for prompt injection

1.5 Non-Goals

1. General-purpose compression: Optimized specifically for LLM API payloads
2. Encryption: Transport security (TLS) is assumed
3. Authentication: Delegated to transport layer
4. Binary protocols: JSON-based wire format only

1.6 Relationship to Other Protocols

Protocol	Relationship
HTTP/1.1, HTTP/2, HTTP/3	M2M operates over HTTP as transport
QUIC (RFC 9000)	Preferred transport for agent-to-agent communication
TLS 1.2+ / TLS 1.3	Required for transport security (built-in with QUIC)
JSON (RFC 8259)	Wire format is JSON-compatible
OpenAI API	Primary target for compression
SSE	Streaming responses supported

1.7 Document Structure

Section	Contents
01-terminology	Definitions and RFC 2119 keywords
02-wire-format	Message structure and encoding
03-message-types	HELLO, ACCEPT, DATA, etc.
04-compression	Algorithms and mappings
05-session-management	State machine and lifecycle
06-security	Threat model and mitigations